Amazon names 34 researchers for AWS Trainium Responsible AI awards

The recipients represent 30 universities and will receive compute credits for projects spanning AI safety, multilingual models, synthetic data and efficient AI systems

Amazon selected 34 researchers from 30 universities for its Fall 2025 Build on Trainium: Responsible AI awards.

Amazon has named 34 researchers from 30 universities as recipients of its Fall 2025 Build on Trainium: Responsible AI awards, providing AWS Trainium compute credits and technical resources for university AI research.

The awards cover five priority areas: AI safety and alignment, multilingual language models, representation engineering, sustainability and small language models, and deep learning models for synthetic data generation.

The awards sit within Build on Trainium, which Amazon describes as a $110 million credit program supporting AI research and university education. That figure applies to the wider initiative. Amazon does not disclose the value of credits assigned to this award cycle or to individual recipients.

According to the company, proposals were assessed for the quality of their scientific content and their potential impact on the research community and society. Recipients gain access to more than 700 Amazon public datasets, AWS AI and machine learning services through Promotional Credits, an Amazon research contact, tutorials and hands-on sessions.

Projects target safety, privacy and multilingual AI

Several projects focus directly on the behavior and reliability of AI models. Shen Shen at the Massachusetts Institute of Technology will research safety benchmarking for AI agent tool use and MCP-specific LoRA mitigations, while Jun Wu at Michigan State University will investigate the safety implications of scaling large language models.

At the University of Illinois at Urbana-Champaign, Yang Wang’s project focuses on protecting young people using multimodal generative AI. Yue Wang at the University of Central Florida will examine game-theoretic approaches to pluralistic AI alignment, and Cheng Tan at Northeastern University will work on reliable and trustworthy large language model services.

Privacy and synthetic data are another strand. Viveck Cadambe at the Georgia Institute of Technology and Haewon Jeong at the University of California Santa Barbara are listed against research into public-private mixtures for differentially private synthetic data generation. Xiaokui Xiao at the National University of Singapore will study Trainium-accelerated synthesis of hierarchical relational data.

Multilingual projects include Songtao Lu’s work at The Chinese University of Hong Kong on multilevel and multiobjective alignment, Yao Lu’s research at University College London into synthetic data for low-resource language model pretraining, and Ryan Shi’s work at the University of Pittsburgh on multilingual large language models used with Indic-language healthcare dialogues.

Elsewhere, Naichen Shi at Northwestern University will investigate large language model hallucination detection and mitigation. Min Xu at Carnegie Mellon University will study language-grounded interpretability for vision transformers and 3D models, while Marios Kogias at Imperial College London will work on deterministic model inference.

Amazon lists project titles and named recipients, but does not provide research timelines, individual credit allocations or expected deliverables for the selected proposals.

Researchers gain access to dedicated Trainium infrastructure

The program is built around AWS Trainium, Amazon’s AI chip for model development. AWS AI Principal Applied Scientist Yida Wang points to work at the University of Illinois Urbana-Champaign as an example of the computing scale involved.

“By leveraging the support from Build on Trainium, University of Illinois Urbana-Champaign researchers are studying topology-aware parallelization strategies for large-scale mixture-of-experts models with as many as one trillion parameters on up to 1,024 Trainium chips,” Wang says.

Amazon says it has created a dedicated research cluster with up to 40,000 Trainium chips. The infrastructure will be available through Amazon EC2 Trn1 instances connected using Amazon EC2 UltraClusters, with research teams and students accessing capacity through self-managed Amazon EC2 Capacity Blocks for machine learning.

Researchers can also use the Neuron Kernel Interface, a Python-based programming environment that provides direct access to hardware primitives and instructions on Trainium and Inferentia chips. Amazon says its research calls welcome projects using open-source machine learning libraries and contributing resources to the wider developer community.

The company plans to conduct multiple Amazon Research Awards proposal rounds under Build on Trainium, although it has not provided a deadline for the next call.